Where Should NGOs Use AI in Communications, and Where Should They Not?

The AI conversation in civil society is stuck between two extremes. On one side are organisations saying it solves everything, on the other those saying they will not touch it. Both produce the same outcome: nobody sets a clear rule, everyone experiments alone and says nothing.

This article proposes a set of rules. Where it genuinely saves time, where it costs you trust, and how to handle the grey area in between.

First, make this distinction

You can use AI for three different jobs, and the risk in each is very different:

In the first row AI is genuinely good today. In the second it helps but review is mandatory. The third is where civil society should not go: in a rights-based organisation, decisions about people must have an explainable justification.

Five things it does well

1. Turning audio and video into text

Field interviews, workshop recordings, panel talks. Transcribing an hour of audio takes three hours by hand and five minutes with AI.

Whisper-based tools are now usable in Turkish. Descript and similar tools handle both transcription and editing. Note: if third parties are audible in the recording, do not upload it to the cloud without their consent.

2. Making long documents readable

A hundred pages of regulatory change, a grant guideline, a monitoring report. "Which clauses in this document concern us" is one of the questions AI answers best.

There are tools built specifically for this. NotebookLM answers only from the documents you upload and shows the source for each statement, which significantly reduces the risk of fabrication. Claude and ChatGPT can also read and summarise long documents.

Still, do not quote a critical clause on the strength of a summary alone. A summary tells you where to look, not what it says.

3. Adapting one message for different audiences

You have a text and you need it written separately for a donor, a beneficiary and a journalist. This is one of the places AI genuinely saves time, because you supply the content and it only changes the register.

The rule: you provide the facts, you ask it for the tone. Do it the other way round and fabrication begins.

4. Translation and language checking

An English grant application, an English web page, an email to an international network. Machine translation has reached the level civil society needs in the past two years.

But watch two things: organisation names and legal terms get mistranslated, and rights-based language disappears. A machine does not know the difference between "refugee" and "asylum seeker", or between person-first and identity-first disability language. Always have someone who knows the field read the translation.

5. Getting started when you are stuck

Blank page anxiety is a real productivity problem. Asking for "ten different headlines on this topic" or "write this opening three different ways" unblocks you.

Do not use the output as it is. Good use looks like this: generate ten options, dislike all of them, notice why you dislike them, and write the eleventh yourself.

Five places not to use it

1. Anything containing personal data. Client details, form submissions, health or legal status. Both under data protection law and ethically, this data does not get pasted into an AI tool. Many free tools may use what you enter for model training.

2. Making decisions about people. Ranking scholarship applications, deciding who is eligible for support, screening job candidates. A rights-based organisation's decision must have an explainable justification, and "the system ranked it this way" is not one.

3. Producing data and statistics. AI can invent a number that does not exist inside an extremely convincing sentence. There should not be a single figure in your report whose source you have not verified.

4. Testimony and storytelling. Having someone's lived experience rewritten in their voice is on the far side of the line in rights-based work. Editing a narrative is one thing; generating it is another.

5. Writing a crisis statement. Getting the tone right in a crisis is where AI is weakest. A crisis text is also a commitment about what the organisation knows and takes responsibility for at that moment. You cannot have a tool make that commitment.

A three-line policy for your organisation

You do not need a long AI policy. Three lines the team will remember are enough:

1. No personal data goes into any AI tool. No exceptions.
2. A human reads and approves every text published in the organisation's name. AI drafts, a person signs.
3. We disclose it where AI made a substantial contribution. Not needed for translation, transcription and language checks; needed for generated images and text.

Discuss these three at a team meeting and write them on one page. A short policy that is followed beats a long one that is not.

Three things to try today

• Transcribe your last workshop recording and compare the time with doing it by hand
• Upload a grant guideline you follow into NotebookLM and ask which clauses concern your organisation
• Write your organisation's three policy lines and share them with the team

Sources

• Think with Google, How brands and agencies use AI in Turkey
• Turkish Personal Data Protection Authority, kvkk.gov.tr, legislation and guidance on personal data processing

If you would like an AI policy or team training for your organisation, get in touch.